Model Overview

ichi234/exp002_stage2_s2_db_merged is a 7.6 billion parameter language model derived from Qwen/Qwen2.5-7B-Instruct. It has been specifically fine-tuned to excel in advanced competitive environments, particularly ALFWorld and DBBench tasks.

Key Capabilities

• ALFWorld Optimization: Stabilizes THOUGHT + ACTION two-line output format and maintains a high rate of legal actions within ALFWorld environments.
• DBBench Enhancement: Improves the stability of Action: Operation / Action: Answer formats and enhances the consistency of generated SQL queries and answers for DBBench.
• Multi-stage Fine-tuning: Utilizes a multi-phase LoRA (bfloat16) training strategy, with varying learning rates and epochs, to progressively refine performance across different aspects of the target tasks.
• Data Augmentation: Incorporates offline distillation using openai/gpt-oss-120b to expand and enhance the training dataset, contributing to improved model robustness and accuracy.

Training Methodology

The model underwent a two-stage training process. Stage 1 focused on stabilizing ALFWorld output formats, followed by DBBench learning, using multiple LoRA phases with a maximum sequence length of 2048. Stage 2 leveraged data augmented via distillation from openai/gpt-oss-120b, training with a maximum sequence length of 4096 to further refine its specialized capabilities.

Good for

• Developing agents for complex interactive environments like ALFWorld.
• Applications requiring precise SQL generation and answer consistency in database interaction tasks (DBBench).
• Research into multi-task learning and agentic behavior in structured environments.